Feel free to start a new discussion here~

[Yang-Xijie]

[Brent-Zoomers]

Hey @Yang-Xijie,

Thanks for the nice work. I have a couple of questions that I don't think are mentioned in the paper unless I missed it.

1. While rendering multiple objects, do you take into account if an object is occluded by another, or do you do your method for every object in the scene?

2. I'm a bit confused by what is meant by Selected Gaussians in Table 4. Is it the percentage of the original Gaussians that is kept or is it the percentage of Gaussians for which your prediction is correct?

Thanks again for the great work and I'm looking forward to a response,
BZ

[Yang-Xijie]

Thanks for your attention!

1

Occlusion between objects is not taken into account during the online selection stage. Clusters or 3D Gaussians are selected independently for each object.

2

Throughout our paper, to derive the percentage, we compare the number of 3D Gaussians selected by our method to that in the original 3DGS asset.

Specifically, suppose the original asset contains $N$ 3D Gaussians. In the offline build stage, we generate multiple levels of detail: $L_0$ ($\sim N$ 3D Gaussians directly taken from original 3DGS asset), $L_1$ ($\sim N/2$ newly generated), $L_2$ ($\sim N/4$ newly generated), ..., $L_k$. During the online selection stage, our method may select $n_0$ 3D Gaussians from $L_0$, $n_1$ from $L_1$, and so on. The selection percentage is calculated as: $(n_0 + n_1 + ... + n_k) / N$.

In Table 4, a percentage close to 100% indicates that our method selects nearly all 3D Gaussians from $L_0$ for the given views from the NVS dataset, showing that it preserves visual quality comparable to the original 3DGS asset when capturing distances are close.